Spatio-Temporal Memory Streaming
Tuesday March 25, 2008
Hamerschlag Hall D-210
Carnegie Mellon University
The memory system remains a key performance bottleneck in modern server
systems. While recent prefetching/streaming proposals have demonstrated
effectiveness at hiding long memory access latencies on certain commercial
workloads, no single technique is effective for online transaction
processing, decision support and web serving.
In this talk, I will present Spatio-Temporal Memory Streaming (STEMS).
STEMS builds upon work in both spatial and temporal memory streaming while
exploiting new observations about repetition within and across spatial
layouts. STEMS dynamically reconstructs a total sequence of predicted
memory accesses by interleaving small-scale spatial predictions into a
longer, recorded sequence that spans distinct regions of memory. Over our
suite of commercial workloads, STEMS achieves similar or higher prediction
coverage than either spatial or temporal memory streaming alone, while
requiring less predictor storage than temporal memory streaming.
Stephen Somogyi is a Ph.D. candidate in Electrical and Computer
Engineering at Carnegie Mellon University, working with Prof. Babak
Falsafi. His research interests focus on memory streaming techniques to
improve the performance of future computer systems. He will graduate later